AI Disruption

AI Disruption

Share this post

AI Disruption
AI Disruption
Alibaba Releases Next-Generation MoE-based Qwen2.5-Max, Surpassing DeepSeek V3

Alibaba Releases Next-Generation MoE-based Qwen2.5-Max, Surpassing DeepSeek V3

Explore Qwen2.5-Max, Alibaba's groundbreaking ultra-large MoE model pushing AI's limits with unmatched performance. Try its multimodal capabilities today!

Meng Li's avatar
Meng Li
Jan 29, 2025
∙ Paid
6

Share this post

AI Disruption
AI Disruption
Alibaba Releases Next-Generation MoE-based Qwen2.5-Max, Surpassing DeepSeek V3
2
Share

"AI Disruption" publication New Year 30% discount link.


The progress of open-source AI is truly insane!

While DeepSeek R1 caused a market earthquake, Alibaba's Tongyi Qianwen dropped another heavy bombshell: Qwen2.5-Max.

Qwen2.5-Max: Exploring the Intelligence of Large-scale MoE Model | Qwen

Regarding the path to AGI, it seems we’ve all been too eager.

For a long time, the common belief was that by scaling up models and data, we’d get closer to AGI. But the reality is that we actually know very little about training ultra-large models.

Today’s Qwen2.5-Max represents an exploration into ultra-large MoE models.

This model, based on the MoE architecture, undergoes massive pretraining with huge datasets, followed by selective SFT and RLHF methods for fine-tuning, using over 20 trillion tokens of training data.

Looking at the data, it's impressive. The arena-hard score is 89.4, ahead of DeepSeek V3 at 85.5 and GPT-4o at 77.9. It performs excellently on tests like MMLU-Pro and GPQA-Diamond.

It surpasses DeepSeek V3 across multiple key metrics, including Arena-Hard and LiveBench, and competes directly with top closed-source models.

This post is for paid subscribers

Already a paid subscriber? Sign in
© 2025 Meng Li
Privacy ∙ Terms ∙ Collection notice
Start writingGet the app
Substack is the home for great culture

Share